R Notebook

source('../settings/settings.R')
source('commonFunctions.R')
library(nlme)
library(lme4)

set.seed(43)
drive1 <- read.csv('../data/processed/analysis/TT1_Drive_1_PP.csv')
drive2 <- read.csv('../data/processed/Analysis/TT1_Drive_2_PP.csv')
drive3 <- read.csv('../data/processed/Analysis/TT1_Drive_3_PP.csv')
drive4 <- read.csv('../data/processed/Analysis/TT1_Drive_4_PP.csv', stringsAsFactors = T)

dfSeg <- data.frame(rep(1, nrow(drive4)), rep(2, nrow(drive4)), rep(3, nrow(drive4)), rep(4, nrow(drive4)))
names(dfSeg) <- c("Seg1", "Seg2", "Seg3", "Seg4")

combinedDf_Seg1 <- cbind(drive4, 
                    drive1$MeanPP_Seg0, 
                    drive2$MeanPP_Seg1, drive3$MeanPP_Seg1, 
                    drive2$MeanPP_Seg0_1, drive3$MeanPP_Seg0_1,
                    drive2$StdPP_Seg1, drive3$StdPP_Seg1,
                    drive2$StdPP_Seg0_1, drive3$StdPP_Seg0_1,
                    drive2$MeanPP_AccHigh1, drive3$MeanPP_AccHigh1,
                    drive2$X.MeanPP_AccLow1, drive3$X.MeanPP_AccLow1,
                    drive2$StdPP_AccHigh1, drive3$StdPP_AccHigh1,
                    drive2$StdPP_AccLow1, drive3$StdPP_AccLow1,
                    dfSeg$Seg1
                  )
combinedDf_Seg2 <- cbind(drive4, 
                    drive1$MeanPP_Seg0, 
                    drive2$MeanPP_Seg2, drive3$MeanPP_Seg2, 
                    drive2$MeanPP_Seg0_2, drive3$MeanPP_Seg0_2,
                    drive2$StdPP_Seg2, drive3$StdPP_Seg2,
                    drive2$StdPP_Seg0_2, drive3$StdPP_Seg0_2,
                    drive2$MeanPP_AccHigh2, drive3$MeanPP_AccHigh2,
                    drive2$X.MeanPP_AccLow2, drive3$X.MeanPP_AccLow2,
                    drive2$StdPP_AccHigh2, drive3$StdPP_AccHigh2,
                    drive2$StdPP_AccLow2, drive3$StdPP_AccLow2,
                    dfSeg$Seg2
                  )
combinedDf_Seg3 <- cbind(drive4, 
                    drive1$MeanPP_Seg0, 
                    drive2$MeanPP_Seg3, drive3$MeanPP_Seg3, 
                    drive2$MeanPP_Seg0_3, drive3$MeanPP_Seg0_3,
                    drive2$StdPP_Seg3, drive3$StdPP_Seg3,
                    drive2$StdPP_Seg0_3, drive3$StdPP_Seg0_3,
                    drive2$MeanPP_AccHigh3, drive3$MeanPP_AccHigh3,
                    drive2$X.MeanPP_AccLow3, drive3$X.MeanPP_AccLow3,
                    drive2$StdPP_AccHigh3, drive3$StdPP_AccHigh3,
                    drive2$StdPP_AccLow3, drive3$StdPP_AccLow3,
                    dfSeg$Seg3
                  )
combinedDf_Seg4 <- cbind(drive4, 
                    drive1$MeanPP_Seg0, 
                    drive2$MeanPP_Seg4, drive3$MeanPP_Seg4, 
                    drive2$MeanPP_Seg0_4, drive3$MeanPP_Seg0_4,
                    drive2$StdPP_Seg4, drive3$StdPP_Seg4,
                    drive2$StdPP_Seg0_4, drive3$StdPP_Seg0_4,
                    drive2$MeanPP_AccHigh4, drive3$MeanPP_AccHigh4,
                    drive2$X.MeanPP_AccLow4, drive3$X.MeanPP_AccLow4,
                    drive2$StdPP_AccHigh4, drive3$StdPP_AccHigh4,
                    drive2$StdPP_AccLow4, drive3$StdPP_AccLow4,
                    dfSeg$Seg4
                  )

common_names <- c("PP_Dev_1_Turning",
                  "PP_Dev_2_Straight", "PP_Dev_3_Straight", 
                  "PP_Dev_2_Turning", "PP_Dev_3_Turning", 
                  "Std_PP_2_Straight", "Std_PP_3_Straight", 
                  "Std_PP_2_Turning", "Std_PP_3_Turning",
                  
                  "Mean_PP_2_AccHigh", "Mean_PP_3_AccHigh",
                  "Mean_PP_2_AccLow", "Mean_PP_3_AccLow",
                  "Std_PP_2_AccHigh", "Std_PP_3_AccHigh",
                  "Std_PP_2_AccLow", "Std_PP_3_AccLow",
                  
                  "Segment")

names(combinedDf_Seg1) <- c(names(drive4), common_names)
names(combinedDf_Seg2) <- c(names(drive4), common_names)
names(combinedDf_Seg3) <- c(names(drive4), common_names)
names(combinedDf_Seg4) <- c(names(drive4), common_names)

# combinedDf_Seg1$Subject <- paste0(as.factor(combinedDf_Seg1$Subject), ".S1")
# combinedDf_Seg2$Subject <- paste0(as.factor(combinedDf_Seg2$Subject), ".S2")
# combinedDf_Seg3$Subject <- paste0(as.factor(combinedDf_Seg3$Subject), ".S3")
# combinedDf_Seg4$Subject <- paste0(as.factor(combinedDf_Seg4$Subject), ".S4")

combinedDf <- rbind(combinedDf_Seg1, combinedDf_Seg2, combinedDf_Seg3, combinedDf_Seg4)

# combinedDf$Subject <- paste0("#", str_pad(combinedDf$Subject, 2, pad="0"))
combinedDf$Segment <- as.factor(combinedDf$Segment)
combinedDf$ActivityEncoded <- factor(ifelse(combinedDf$Activity == "NO", "1", ifelse(combinedDf$Activity == "C", "2", "3")))

combinedDf <- combinedDf[complete.cases(combinedDf),]
combinedDf$Subject = as.factor(combinedDf$Subject)

model = lm(PP_After ~ 
              PP_Dev_2_Straight + 
              PP_Dev_3_Straight +
              PP_Dev_2_Turning + 
              PP_Dev_3_Turning + 
              Std_PP_2_Straight + 
              Std_PP_3_Straight + 
              Std_PP_2_Turning +
              Std_PP_3_Turning +
              # PP_Prior +
              factor(ActivityEncoded),
            data=combinedDf, random = ~1|factor(Subject), method = "REML")

method = 'REML' is not supported. Using 'qr'In lm.fit(x, y, offset = offset, singular.ok = singular.ok, ...) :
 extra argument ‘random’ will be disregarded

# anova(model)
summary(model)

Call:
lm(formula = PP_After ~ PP_Dev_2_Straight + PP_Dev_3_Straight + 
    PP_Dev_2_Turning + PP_Dev_3_Turning + Std_PP_2_Straight + 
    Std_PP_3_Straight + Std_PP_2_Turning + Std_PP_3_Turning + 
    factor(ActivityEncoded), data = combinedDf, method = "REML", 
    random = ~1 | factor(Subject))

Residuals:
     Min       1Q   Median       3Q      Max 
-0.17764 -0.07137  0.00092  0.05144  0.34502 

Coefficients:
                         Estimate Std. Error t value Pr(>|t|)    
(Intercept)              -0.02294    0.04870  -0.471   0.6393    
PP_Dev_2_Straight         0.68712    0.15583   4.410 4.45e-05 ***
PP_Dev_3_Straight        -0.37734    0.23618  -1.598   0.1155    
PP_Dev_2_Turning         -0.22015    0.14683  -1.499   0.1391    
PP_Dev_3_Turning          0.38632    0.22101   1.748   0.0857 .  
Std_PP_2_Straight         0.02144    0.38757   0.055   0.9561    
Std_PP_3_Straight         0.50369    0.36818   1.368   0.1765    
Std_PP_2_Turning         -0.08343    0.50205  -0.166   0.8686    
Std_PP_3_Turning         -0.95800    0.60836  -1.575   0.1207    
factor(ActivityEncoded)2  0.08647    0.03285   2.633   0.0108 *  
factor(ActivityEncoded)3  0.17032    0.03184   5.349 1.51e-06 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 0.0993 on 59 degrees of freedom
Multiple R-squared:  0.5613,    Adjusted R-squared:  0.4869 
F-statistic: 7.548 on 10 and 59 DF,  p-value: 1.338e-07

plot(model)

No Random Effects

linearModel1 <- lm(PP_After ~ 
                Mean_PP_2_AccHigh
              + Mean_PP_2_AccLow
              + Mean_PP_3_AccHigh
              + Mean_PP_3_AccLow
              + Std_PP_2_AccHigh
              + Std_PP_2_AccLow
              + Std_PP_3_AccHigh
              + Std_PP_3_AccLow
              # + PP_Prior
              + factor(ActivityEncoded),
            data=combinedDf)

# anova(model)
summary(linearModel1)

Call:
lm(formula = PP_After ~ Mean_PP_2_AccHigh + Mean_PP_2_AccLow + 
    Mean_PP_3_AccHigh + Mean_PP_3_AccLow + Std_PP_2_AccHigh + 
    Std_PP_2_AccLow + Std_PP_3_AccHigh + Std_PP_3_AccLow + factor(ActivityEncoded), 
    data = combinedDf)

Residuals:
      Min        1Q    Median        3Q       Max 
-0.143483 -0.066038 -0.007552  0.051595  0.300313 

Coefficients:
                         Estimate Std. Error t value Pr(>|t|)    
(Intercept)              -0.06552    0.03640  -1.800  0.07695 .  
Mean_PP_2_AccHigh         1.63376    0.50590   3.229  0.00203 ** 
Mean_PP_2_AccLow         -1.15253    0.49763  -2.316  0.02405 *  
Mean_PP_3_AccHigh         0.74436    0.31520   2.362  0.02152 *  
Mean_PP_3_AccLow         -0.66142    0.33913  -1.950  0.05589 .  
Std_PP_2_AccHigh         -1.28974    1.45755  -0.885  0.37982    
Std_PP_2_AccLow           0.79583    1.15728   0.688  0.49435    
Std_PP_3_AccHigh          0.15278    0.94374   0.162  0.87195    
Std_PP_3_AccLow           0.91693    0.83810   1.094  0.27838    
factor(ActivityEncoded)2  0.08724    0.03041   2.869  0.00571 ** 
factor(ActivityEncoded)3  0.14810    0.03043   4.867 8.83e-06 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 0.09353 on 59 degrees of freedom
Multiple R-squared:  0.6108,    Adjusted R-squared:  0.5448 
F-statistic: 9.259 on 10 and 59 DF,  p-value: 5.37e-09

plot(linearModel1)

With Random Effects

linearModel1 <- lme(PP_After ~ 
                Mean_PP_2_AccHigh
              + Mean_PP_2_AccLow
              + Mean_PP_3_AccHigh
              + Mean_PP_3_AccLow
              + Std_PP_2_AccHigh
              + Std_PP_2_AccLow
              + Std_PP_3_AccHigh
              + Std_PP_3_AccLow,
              # + factor(ActivityEncoded),
              random=~1|factor(Subject),
            data=combinedDf)

# anova(model)
summary(linearModel1)

Linear mixed-effects model fit by REML
 Data: combinedDf 

Random effects:
 Formula: ~1 | factor(Subject)
        (Intercept)     Residual
StdDev:   0.1466446 2.557679e-17

Fixed effects: PP_After ~ Mean_PP_2_AccHigh + Mean_PP_2_AccLow + Mean_PP_3_AccHigh +      Mean_PP_3_AccLow + Std_PP_2_AccHigh + Std_PP_2_AccLow + Std_PP_3_AccHigh +      Std_PP_3_AccLow 

 Correlation: 
                  (Intr) M_PP_2_AH M_PP_2_AL M_PP_3_AH M_PP_3_AL S_PP_2_AH S_PP_2_AL S_PP_3_AH
Mean_PP_2_AccHigh -0.173                                                                      
Mean_PP_2_AccLow   0.172 -0.961                                                               
Mean_PP_3_AccHigh  0.035 -0.001    -0.023                                                     
Mean_PP_3_AccLow   0.048 -0.062     0.023    -0.694                                           
Std_PP_2_AccHigh  -0.061  0.155    -0.193     0.086    -0.076                                 
Std_PP_2_AccLow   -0.037  0.036    -0.019    -0.132     0.087    -0.873                       
Std_PP_3_AccHigh  -0.025  0.066    -0.117    -0.454     0.310    -0.336     0.191             
Std_PP_3_AccLow    0.060  0.191    -0.208     0.377    -0.323     0.302    -0.238    -0.528   

Standardized Within-Group Residuals:
       Min         Q1        Med         Q3        Max 
-5.4259302 -0.5425930  0.0000000  0.8138895  4.3407442 

Number of Observations: 70
Number of Groups: 21 

plot(linearModel1)

linearModel1 <- lm(PP_After ~ 
                Mean_PP_2_AccHigh
              + Mean_PP_2_AccLow
              + Mean_PP_3_AccHigh
              + Mean_PP_3_AccLow
              # + PP_Prior
              + factor(ActivityEncoded), 
            data=combinedDf)

# anova(model)
summary(linearModel1)

Call:
lm(formula = PP_After ~ Mean_PP_2_AccHigh + Mean_PP_2_AccLow + 
    Mean_PP_3_AccHigh + Mean_PP_3_AccLow + factor(ActivityEncoded), 
    data = combinedDf)

Residuals:
     Min       1Q   Median       3Q      Max 
-0.14161 -0.06916 -0.00742  0.04809  0.30760 

Coefficients:
                         Estimate Std. Error t value Pr(>|t|)    
(Intercept)              -0.05248    0.02751  -1.907  0.06103 .  
Mean_PP_2_AccHigh         1.64939    0.38841   4.246 7.28e-05 ***
Mean_PP_2_AccLow         -1.14578    0.37518  -3.054  0.00331 ** 
Mean_PP_3_AccHigh         0.65161    0.26162   2.491  0.01540 *  
Mean_PP_3_AccLow         -0.60534    0.28421  -2.130  0.03709 *  
factor(ActivityEncoded)2  0.08906    0.02976   2.993  0.00394 ** 
factor(ActivityEncoded)3  0.15955    0.02892   5.516 6.92e-07 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 0.09303 on 63 degrees of freedom
Multiple R-squared:  0.5888,    Adjusted R-squared:  0.5496 
F-statistic: 15.03 on 6 and 63 DF,  p-value: 1.372e-10

plot(linearModel1)

Machine Learning

ppAfter <- combinedDf$PP_After
ppAfterArray <- matrix(ppAfter, nrow = 1,ncol = length(ppAfter))
  
thresholdPPAfter <- otsu(ppAfterArray, range=c(min(ppAfter), max(ppAfter))) # Expected Threshold > 0.10123
print(paste0('Threshold: ', thresholdPPAfter))

[1] "Threshold: 0.101235546875"

selectedDf <- combinedDf %>% select(
                  "Subject", "Activity", "PP_After", # "PP_Prior",
                  "Mean_PP_2_AccHigh", "Mean_PP_3_AccHigh",
                  "Mean_PP_2_AccLow", "Mean_PP_3_AccLow",
                  "Std_PP_2_AccHigh", "Std_PP_3_AccHigh",
                  "Std_PP_2_AccLow", "Std_PP_3_AccLow")

selectedDf$Subject <- NULL
selectedDf$Activity_NO <- ifelse(selectedDf$Activity == "NO", 1, 0)
selectedDf$Activity_C <- ifelse(selectedDf$Activity == "C", 1, 0)
selectedDf$Activity_M <- ifelse(selectedDf$Activity == "M", 1, 0)
selectedDf$Activity <- NULL

# selectedDf$PP_Dev_1_Turning <- NULL
# selectedDf$Std_PP_2_Straight <- NULL
# selectedDf$Std_PP_2_Turning <- NULL
# selectedDf$Std_PP_3_Straight <- NULL
# selectedDf$Std_PP_3_Turning <- NULL
# 
# # According to Linear model
# selectedDf$PP_Dev_2_Straight <- abs(selectedDf$PP_Dev_2_Straight)
# selectedDf$PP_Dev_3_Straight <- abs(selectedDf$PP_Dev_3_Straight)
# selectedDf$PP_Dev_2_Turning <- abs(selectedDf$PP_Dev_2_Turning)
# selectedDf$PP_Dev_3_Turning <- abs(selectedDf$PP_Dev_3_Turning)
# selectedDf$PP_Prior <- abs(selectedDf$PP_Prior) # NULL

selectedDf$Class <- ifelse(selectedDf$PP_After >= thresholdPPAfter, T, F)
selectedDf$PP_After <- NULL

print(names(selectedDf))

 [1] "Mean_PP_2_AccHigh" "Mean_PP_3_AccHigh" "Mean_PP_2_AccLow"  "Mean_PP_3_AccLow"  "Std_PP_2_AccHigh"  "Std_PP_3_AccHigh" 
 [7] "Std_PP_2_AccLow"   "Std_PP_3_AccLow"   "Activity_NO"       "Activity_C"        "Activity_M"        "Class"            

# library(mefa)
# combinedDf <- rep(combinedDf, 10) 

set.seed(39)
n_folds <- 10
params <- param <- list(objective       = "binary:logistic", 
               booster          = "gbtree",
               eval_metric      = "auc",
               eta              = 0.1,
               max_depth        = 10,
               alpha            = 1,
               lambda           = 0,
               gamma            = 0.45,
               min_child_weight = 0.3,
               subsample        = 1,
               colsample_bytree = 1)
           
# XGBoost Model         
xgb_m <- xgb.cv(   params               = param,
                  data = as.matrix(selectedDf %>% select(-Class)) ,
                  label =  selectedDf$Class,
                  nrounds             = 100,
                  verbose             = F,
                  prediction          = T,
                  maximize            = F, # Change this value to F will help to run with more itineration
                  nfold               = n_folds,
                  metrics             = c("auc", "error"),
                  early_stopping_rounds = 50,
                  stratified            = T,
                  scale_pos_weight      = 1)

# xgb_m$evaluation_log[xgb_m$best_iteration,"test_auc_mean"]
xgb_m$evaluation_log[xgb_m$best_iteration,]

NA

Performance Metrics

# Prediction
selectedDf$clsPred <- round(xgb_m$pred)

computePerformanceResults <- function(sdat){
  sdat = sdat[complete.cases(sdat),]
  acc = sum(sdat[,1] == sdat[,2])/nrow(sdat)
  conf_mat = table(sdat)
  specif = conf_mat[1,1]/sum(conf_mat[,1])
  sensiv = conf_mat[2,2]/sum(conf_mat[,2])
  preci =  conf_mat[2,2]/sum(conf_mat[2,])
  npv =    conf_mat[1,1]/sum(conf_mat[1,])
  return(c(acc,specif,sensiv,preci,npv))
}

# Get average performance
performance <- computePerformanceResults(selectedDf %>% select(Class, clsPred))
acc <- performance[1]
prec <- performance[4]
recall <- performance[3]
spec <- performance[2]
npv <- performance[5]
f1 <- (2 * recall * prec) / (recall + prec)
auc <- as.numeric(xgb_m$evaluation_log[xgb_m$best_iteration, "test_auc_mean"])

print(paste("Accuracy=", round(acc, 2)))

[1] "Accuracy= 0.81"

print(paste("Precision=", round(prec, 2)))

[1] "Precision= 0.7"

print(paste("Recall=", round(recall, 2)))

[1] "Recall= 0.84"

print(paste("Specificity=", round(spec, 2)))

[1] "Specificity= 0.8"

print(paste("NPV=", round(npv, 2)))

[1] "NPV= 0.9"

print(paste("F1=", round(f1, 2)))

[1] "F1= 0.76"

print(paste("AUC=", round(auc, 2)))

[1] "AUC= 0.9"

# Importance
bst <- xgboost(   params               = param,
                  data = as.matrix(selectedDf %>% select(-c(Class, clsPred))) ,
                  label =  selectedDf$Class,
                  nrounds             = 100,
                  verbose             = F,
                  prediction          = T,
                  maximize            = F, # Change this value to F will help to run with more itineration
                  nfold               = n_folds,
                  metrics             = c("auc", "error"),
                  early_stopping_rounds = 50,
                  stratified            = T,
                  scale_pos_weight      = 1)
importanceDf <- xgb.importance(colnames(selectedDf %>% select(-c(Class, clsPred))), model = bst)
print(importanceDf)

library(pROC)

dfROC <- pROC::roc(response = ifelse(selectedDf$Class==T, 1, 0),
               predictor = round(xgb_m$pred),
               levels=c(0, 1), direction = "<")

# it = which.max(xgb_m$evaluation_log$test_auc_mean)
# best.iter = xgb_m$evaluation_log$iter[it]
# best.iter 

plot(pROC::roc(response = ifelse(selectedDf$Class==T, 1, 0),
               predictor = round(xgb_m$pred),
               levels=c(0, 1), direction = "<"), 
     legacy.axes = TRUE,
     main="ROC Curve", 
     lwd=1.5) 

Plot feature importance

yAxis <- list(
  title = 'Importance',
  range=c(0.0, 1.0)
)
xAxis <- list(
  title = ''
)

importanceDf$Feature <- factor(importanceDf$Feature, levels = importanceDf[order(-Gain),]$Feature)
fig_Importance <- plot_ly(importanceDf, x = ~Feature, y = ~Gain, type = 'bar', name = 'Gain', width=600) %>%
  add_trace(y = ~Cover, name = 'Cover') %>% 
  add_trace(y = ~Frequency, name = 'Frequency') %>% 
  layout(yaxis = yAxis, xaxis=xAxis, barmode = 'group', title="Feature Importance") %>% 
  config(.Last.value, mathjax = 'cdn')

htmltools::tagList(fig_Importance)